For today’s post, we reached out to Roberto Barriga, Zigurat’s professor of Global MBA in Digital Business, to ask him to share his recommendations for companies that are interested in applying Advanced Analytics and a Machine Learning pilot in their manufacturing process.
Manufacturers have traditionally been very successful using data to increase efficiency and quality, but are finding that lean production and cost-cutting are no longer enough to remain competitive. The goal today is to integrate and gain insights from data across their complex global and often fragmented supply chains.
Manufacturers generate and store data from many sources across the supply chain, including process control instruments, supply chain management systems, and systems that monitor the performance of products after they’ve been sold. Being able to access hidden data and integrate all of this data across multiple sources provides valuable insights and a competitive advantage.
These insights can lead to improvements in design and production, product quality, forecasting, more targeted products, distribution, and identify hidden bottlenecks in the production process. The following text describes some recommendations a company may follow when choosing an Advanced Analytics technique and applying a Machine Learning pilot in a manufacturing process.
Stage #1: Data collection
The goal of Machine Learning is to ensure that computers and machines can “learn”, without having to be programmed, to make business decisions in a faster and more optimized way. Machine Learning can be both supervised and unsupervised but, ultimately, data is the main source of study. In this way, the more data a company generates, the more information for the algorithms and more capacity to achieve effective predictive models.
The first step to understanding the problem is to ask yourself: why? Until you understand the root of the problem. Likewise, the aim and the opportunity for improvement to be applied must be established. One example could be to reduce the time that a production line spends on a specific process.
Usually, for collecting data from machinery, the Unified OPC Architecture (OPC UA) is used. It’s a machine-to-machine communication protocol for industrial automation and characterized by its focus on communicating with industrial equipment and systems for data collection and control.
This system allows users to customize how data is organized and how information about that data is reported and then save them in the ‘cloud’ for the algorithm to process them.
Stage #2: Prepare the data
The Machine Learning algorithm needs data, so the data collection phase is the first; as important as understanding the problem is understanding the data that we have available. An exploratory data analysis should be done to become familiar with them.
In an exploratory analysis, graphs, correlations, and descriptive statistics are usually made to better understand what the data is telling us. It also helps to estimate whether the data collected is okay, enough, and relevant, to build a model.
Once you have the data and look for patterns, you continue with the process of cleaning the data. The goal of this stage is to manipulate and convert the data in ways that produce better results.
In this section, eliminate or infer missing data, classify the values of the variables, control the numerical values or scale them so that they can be comparable and useful.
Also, Machine Learning algorithms work much better if relevant characteristics are put in place instead of raw data.
For example, it is much easier for the team to know the temperature in degrees Celsius than to know how much a few milligrams of mercury have expanded in a traditional thermometer: It is very useful to transform the data to make the learning task easier.
Stage #3: Choice of algorithm
During this phase, the first thing to do is choose what type of machine learning technique needs to be used, and thus, the machine learning algorithm will automatically learn to obtain the appropriate results with the historical data that has been prepared.
Once the data has been processed, it is time to choose the most appropriate algorithm in relation to the problem that needs to be solved: This is the section in which you should check whether a supervised learning algorithm or unsupervised learning will be used.
Within the unsupervised learning algorithms, there are options such as K-Means Clustering, Hierarchical Clustering, or Principal Component Analysis.
Within the supervised learning algorithms, options such as Linear Regression, Logistic Regression, Decision Tree Regression, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree Classification, among others can be used.
In this way, having chosen the algorithm, the next step is to separate the data, since a percentage of the data, approximately 70% of the total, will be used as training data and will be the one to which the selected algorithm will be applied. In this activity, the goal is to achieve the aim initially set.
Stage #4: Model integrated into a System
Finally, the error analysis phase requires an effort to analyze errors, it is important to understand what needs to be done to improve the results of machine learning.
In most cases, these options will be:
– Use complex model
– Use a simpler model
– Collect more data and/or more features
– Develop a better understanding of the problem and better understand what the next step is
In the error analysis stage, it is important to monitor that the model is capable, this is, the ability of Machine Learning models to produce good results when using new data.
If a production line process optimization wants to get really good results, you must iterate over the previous phases several times. With each iteration, our understanding of the problem and the data will increase. This will allow us to design better relevant features and reduce generalization error. Almost always, having more data helps since more data and a simple model tend to perform better than a complex model with fewer data.
Once we are satisfied with the error, we must compare it to the error in the current solution. If it is better enough, we will integrate the machine learning model into our manufacturing process.
Machine Learning uses a process that works with a computer algorithm that finds a pattern in the data and predicts the most likely results. Machine Learning patterns are constantly updated when new data is entered, making them increasingly accurate in their predictions the longer they operate.
In terms of corporate advantages, Machine Learning algorithms can be an answer to a variety of manufacturing complexities and can help to improve operations in production plants, as for example, this technology allows a factory to improve and optimize production lines processes, by reducing production times, eliminating human errors, reducing costs, and less energy use.
Would you like to know more about Machine Learning algorithms? Learn about manufacturing in the digital era in our MBA program.
Professor of the Global MBA in Digital Business
Business Partner & Services Delivery MGR at Almirall
Computer and Systems Informatics Department, Universitat Politècnica de València